基于LSTM-压缩感知的居民用户负荷数据异常识别及缺失数据修复融合清洗模型  

LSTM-compressed sensing-based fusion cleaning model for abnormal identification of residential user load data and missing data repair

作　　者：刘雁行 乔如妤 梁楠 徐恺 于凯 李旭东 冯军 孔繁春 李雨攸 温洪林 顾洁[2] LIU Yanxing;QIAO Ruyu;LIANG Nan;XU Kai;YU Kai;LI Xudong;FENG Jun;KONG Fanchun;LI Yuyou;WEN Honglin;GU Jie(Inner Mongolia Power(Group),Power marketing and operation management division,Hohhot 010090,China;Shanghai Jiao Tong University,Department of Electrical Engineering,Shanghai 200240,China)

机构地区：[1]内蒙古电力(集团)有限责任公司电力营销与运营管理分公司,内蒙古自治区呼和浩特010090 [2]上海交通大学电气工程系,上海200240

出　　处：《供用电》2025年第3期104-114,共11页Distribution & Utilization

基　　金：国家自然科学基金项目(52307119);内蒙古电力(集团)有限责任公司2023年第一批科技项目(2023-5-46)。

摘　　要：精准的负荷数据是电力系统规划运行基础,但实践中常见数据缺失和数据异常现象,导致负荷预测误差增大,从而影响规划方案和调度、交易策略的可行性与经济性。针对当前数据处理中异常识别和缺失值修复分别建模,以及连续数据缺失修复效果不佳的问题,提出了一种基于长短期记忆网络(long-short term memory,LSTM)和压缩感知的数据融合清洗模型。该模型利用长短期记忆网络识别负荷异常并构建观测矩阵,结合K-means算法和K-奇异值分解算法训练字典矩阵,通过贪婪追踪算法对负荷序列进行重构,实现了负荷数据异常识别和缺失数据修复融合。以实际居民用户负荷数据为例,对比分析了所提出的清洗模型与多种常用清洗模型的处理效果,验证了模型的有效性。Accurate load data is the basis of power system planning and operation,but data missing and data anomalies are common in practice,leading to increased load forecasting errors,which affects the feasibility and economy of planning schemes and scheduling and trading strategies.In response to the problems of anomaly identification and missing data repair are modeled separately in current data processing,as well as the poor effect of continuous data missing repair,this paper proposes a data fusion cleaning model based on long and short-term memory network and compressed perception.The model identifies load anomalies and constructs observation matrices using the long and short-term memory network,combines the K-means algorithm and the K-singular value decomposition algorithm to train dictionary matrices,and reconstructs the load sequences through the greedy tracking algorithm,which achieves the fusion of load data anomaly identification and missing data repair.Taking the actual residential user load data as an example,the processing effect of the cleaning model proposed in this paper and a variety of commonly used cleaning models are compared and analyzed,and the effectiveness of the model in this paper is verified.

关 键 词：负荷预测 异常识别 缺失值修复 长短期记忆网络 压缩感知 

分 类 号：TM721[电气工程—电力系统及自动化]